Secure AI Deployment and Monitoring

A retrieval question-answering application becomes a different kind of system the day it gets a URL, a customer and a login. This course takes one and deploys it properly, control by control, and proves every control against the running service rather than against a configuration file. You start by mapping the trust boundaries and writing a threat model that names a control for each of the ten OWASP risks for LLM applications. Then you build: FastAPI in a non-root, read-only container on a network with no route out, behind nginx terminating TLS; instruction and data separated in the template, with a request guard for direct injection and an ingest sanitiser for the instructions people write inside documents; an output filter that closes the exfiltration channels an answer can carry; per-client keys hashed at rest and bound to one workspace and one scope; a token bucket, a per-client daily token budget in shared state, a concurrency ceiling and size caps at two layers; one structured log line per request carrying digests instead of text; metrics scraped per instance, alert rules with an evaluator, and an eval canary that runs against the deployment; and a release pipeline that builds once, gates the artefact against a candidate stack, deploys the image id it gated and rolls back to one that passed. Two faults are seeded on purpose so you can see where each stage of a gate earns its place, and a third is seeded on a disposable copy of the stack so you can watch an alert fire and then put it away. You finish with a kill switch that works without a deploy, a runbook a stranger could follow, and a final project that onboards a third workspace, brings the whole stack up from nothing and hands it over. Everything runs offline against a local deterministic stub. No provider SDK is installed, no provider host is reachable from the application network, and at no point does the course ask you for an API key. The course is explicit about what that does and does not prove. This is a learning pathway toward an AI application developer role. It does not promise employment, seniority, salary or any vendor certification; completing it earns an Ultiblob Certificate of Completion.